AI's Next Bottleneck Is Money, Not Chips
Broadcom is lending Anthropic up to $42 billion, and Nvidia wants to build a $500 billion market for loans backed by GPUs. The AI race may turn on what rate buyers pay to borrow.
The chip seller is now the lender
As AI infrastructure grows, the role of chipmakers is changing. In the past, customers raised money and bought chips. Now suppliers are moving inside the customer's financing structure.
According to Anthropic's IPO prospectus, as seen by Reuters, Broadcom has agreed to provide Anthropic with up to $42 billion in financial support. Anthropic has committed to $125.2 billion of TPU compute leases over five years. The financing limit alone could cover about one-third of that.
The relationship does not stop there. Broadcom is Anthropic's chip design partner and hardware supplier, and also its financing partner. Anthropic is expected to become the largest customer of Broadcom's chip design business in 2027. Some of the debt could convert to equity after Anthropic's IPO.
The structure is powerful, but it is also complicated. Anthropic itself warned in the prospectus that Broadcom's dual role as hardware supplier and financing partner could create potential conflicts of interest over access to compute and pricing.
Nvidia wants GPUs to look like aircraft
Nvidia's experiment goes a step further. The company is working with financial firms including Blackstone, Apollo and KKR to build a chip-backed financing market of up to $500 billion. AI developers would borrow against GPUs and long-term customer contracts to secure more compute.
Nvidia sees its AI compute as a productive, durable asset that can generate cash flow for years and can be transferred to other users. The logic is that, like leased aircraft, GPUs can become financeable infrastructure assets.
Wall Street has not put the same price tag on them yet. According to Reuters reporting, banks typically assume a depreciation schedule of three to four years for GPUs. They are more conservative about the earning life of up to 10 years that Nvidia argues for. Some lenders are asking for Nvidia guarantees, long-term contracts from investment-grade customers, and extra repayment protection.
The issue is economic life, not physical life. A five-year-old GPU can still switch on. Whether it can earn money with competitive power efficiency and compute cost is a different question. So lenders care less about how many chips exist and more about the contracted cash flow those chips can produce.
A new gauge: the Cost of AI Capital
Taken together, the two deals suggest a new way to read the AI investment cycle: the Cost of AI Capital.
GPU prices alone are no longer enough. Investors also need to look at the interest rate on the borrowing used to buy the GPU, the share of its value that lenders accept as collateral, how much Nvidia or Broadcom guarantees, and how many years of contract the end customer provides.
Two data centers with identical performance can have very different economics if one operator's cost of capital is 6% and the other's is 12%. If compute lease rates are the same, the difference in interest expense becomes a difference in profitability. Balance sheets and access to financing could become a new moat in AI, alongside model quality and GPU supply.
How the AI capital market is structured
- Chip suppliers: Broadcom, Nvidia
- Financing: vendor financing, private credit
- AI customers: AI labs, data centers
- Cash flow: compute leases, AI revenue
Good leverage, or circular financing?
There is no reason to call this structure a bubble outright. AI has real usage and revenue. Cases already exist where contracted cash flow backs the debt, such as CoreWeave's $8.5 billion GPU-backed loan supported by Meta's contracts.
Still, investors need to watch how the money circulates. A supplier funds a customer, the customer uses that money to buy the supplier's chips and compute, and the supplier then expands financing on the strength of that revenue. The larger that loop grows, the more the quality of final demand matters.
In the end, the question is not how big AI capex is. It is whether the cash flow from the final AI services built on that infrastructure can cover interest and principal. If the answer weakens, vendor financing could turn from a growth accelerant into a device for adding leverage.
Insight Times Editorial Desk
